---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': A
          '1': B
          '2': C
          '3': D
          '4': E
          '5': F
          '6': G
          '7': H
          '8': I
          '9': J
  splits:
  - name: train
    num_bytes: 6842235.510231657
    num_examples: 14979
  - name: test
    num_bytes: 1715013.5296924065
    num_examples: 3745
  download_size: 8865158
  dataset_size: 8557249.039924063
task_categories:
- image-classification
- image-to-image
- text-to-image
- image-to-text
tags:
- mnist
- notmnist
pretty_name: notMNIST
size_categories:
- 10K<n<100K
---
# Dataset Card for "notMNIST"

## Overview

The notMNIST dataset is a collection of images of letters from A to J in various fonts. It is designed as a more challenging alternative to the traditional MNIST dataset, which consists of handwritten digits. The notMNIST dataset is commonly used in machine learning and computer vision tasks for character recognition.

## Dataset Information

  ```lua
  Number of Classes: 10 (A to J)
  Number of Samples: 187,24
  Image Size: 28 x 28 pixels
  Color Channels: Grayscale

## Dataset Structure

The dataset is split into a training set and a test set. Each class has its own subdirectory containing images of that class. The directory structure is as follows:

  ```lua
  notMNIST/
  |-- train/
  |   |-- A/
  |   |-- B/
  |   |-- ...
  |   |-- J/
  |
  |-- test/
  |   |-- A/
  |   |-- B/
  |   |-- ...
  |   |-- J/


## Acknowledgements

  http://yaroslavvb.blogspot.com/2011/09/notmnist-dataset.html
  https://www.kaggle.com/datasets/lubaroli/notmnist


## Inspiration

This is a pretty good dataset to train classifiers! According to Yaroslav:

  > Judging by the examples, one would expect this to be a harder task
  than MNIST. This seems to be the case -- logistic regression on top of
  stacked auto-encoder with fine-tuning gets about 89% accuracy whereas
  same approach gives got 98% on MNIST. Dataset consists of small
  hand-cleaned part, about 19k instances, and large uncleaned dataset,
  500k instances. Two parts have approximately 0.5% and 6.5% label error
  rate. I got this by looking through glyphs and counting how often my
  guess of the letter didn't match it's unicode value in the font file.